Development of a controller for robotic manipulation through Learning from demonstration

This thesis proposes a Learning from Demonstration (LfD) approach designed to generalize and extract relevant features of desired motion trajectories for robotic manipulation tasks, with the specific objective of learning a sliding and picking task exploiting environmental constraints and force sensor data. Learning from Demonstration is a powerful approach in robotics, since robots can acquire new skills by observing, modeling and imitating human demonstrations of a task. This method leverages human expertise to teach robots complex movements, reducing the need for explicit programming. Research in the field of Lfd is facilitating, even for non-expert users, to teach new tasks to robots with few demonstrations, enabling robotics and skills learning to be used in a variety of fields of applications and dynamic environments. The developed method is based on probabilistic modeling, specifically a mixture of Hidden Markov Model(HMM) and Gaussian Mixture Model(GMM). Pose data and force data are processed,cleaned and aligned in time through Dinamic Time Warping. A crucial role in the method is represented by forces and torques data sensed by the force sensor during the contact with the environment, through which the movement is divided in three primitives. After this division, a continuous HMM model is trained to be able to switch between primitives and enabling a high level control of the trajectory. At a lower level a GMM for each primitive is trained to model the pose of the end effector and predicting it, giving time as input, with Gaussian Mixture Regression.Both HMM and GMM models are tested on training and test set in order to fine tune parameters, such as the number of gaussians distribution. The proposed approach, combining these two models, is implemented and validated through simulations and real-world experiments, showing good performances in predicting primitives and generalizing the trajectory. The main advantage of this algorithm is its ability to generalize from few demonstrations, resulting in high-quality motion reproduction. The probabilistic approach enables modeling of complex trajectories with a limited number of demonstrations, leveraging force sensor data and environmental constraints to enhance robustness.